Do Memory Types Earn Their Keep?
I built a typed memory system, so when a leading minimal baseline in this year's ICLR submissions asks whether all that structure is worth it, I have to take the question personally. This is me taking it personally, and then answering it with the rest of the literature.
TL;DR. A few weeks ago I published an essay arguing there are five distinct types of AI memory. Then I read through the memory papers submitted to ICLR 2026 and found the strongest possible rebuttal sitting near the top of my own reading list: ENGRAM, a deliberately minimal, lightweight memory design that asks whether the elaborate taxonomies some of us have built actually earn their complexity. It is a fair question and I refuse to strawman it. But the same crop of submissions contains the counter-case: work testing whether memory types need distinct storage and retrieval strategies, psych-inspired architectures built around evolving schemas, and an RL taxonomy with explicit evaluation methodology. My read: the burden of proof sits on structure, the new work finally makes that burden testable, and the honest way to hold a position like mine is to keep naming what would change my mind.
Every builder has a paper they are afraid of. Not a rival product, a paper. The one that, if its result holds up and generalizes, means you overbuilt. For me, this year, that paper is called ENGRAM, an ICLR 2026 submission whose pitch is effective, lightweight memory orchestration for conversational agents. Lightweight is the operative word. It is a minimal typed design, and its existence poses a question directly at systems like the one I build: if a small, simple memory layer gets you most of the way, what exactly is all your machinery for?
I want to answer that question in public, because I think the answer is genuinely interesting, and because the alternative, quietly hoping nobody asks it, is how you end up believing your own architecture diagrams.
The Skeptic's Case, Stated Properly
The minimalist position is not stupid, and it is not new. It is the memory-flavored version of the bitter lesson: every time researchers encode human-designed structure into a system, a simpler approach with more compute eventually embarrasses them. Under that view, a memory system should be a thin orchestration layer. Store what the agent saw, retrieve what looks relevant, let the model do the cognition. Types, graphs, lifecycles, consolidation stages, all of it is premature ontology, complexity that will age like milk.
There is also a nastier, more practical version of the argument. Complexity has a carrying cost. Every distinct memory type is a distinct code path, a distinct failure mode, a distinct thing to tune. If a minimal baseline scores within a few points of an elaborate system on the benchmarks we currently have, then the elaborate system is paying rent it cannot afford. That is the question ENGRAM forces, and I think anyone building typed memory owes it a straight answer rather than an eye roll.
Where the Skeptic Is Right
Two concessions, because they are true.
First, most memory retrievals are boring. The majority of what an assistant needs on a given turn is recency and similarity, and a flat store with good ranking serves that fine. If you evaluate on workloads dominated by boring retrievals, minimal designs will look great, because for that slice they are great.
Second, taxonomy is cheap to invent and expensive to validate. Anyone can propose seven types of memory in a blog post. The field has produced plenty of cognitive-sounding architectures whose categories exist because they sounded right, not because ablations demanded them. Skepticism toward typed memory is partly earned by the people who typed first and measured never.
So the burden of proof sits on structure. I accept that placement. Here is the evidence that makes me think structure has earned a harder look.
The Case for Types
The most direct entry is an ICLR 2026 submission titled Memory Type Matters, which tests exactly the claim under dispute: whether long-term memory improves when different memory types get distinct storage and retrieval strategies rather than one uniform pipeline. That is the whole architectural bet of typed memory, isolated and put under test: do distinct strategies for different memory types justify themselves against a uniform approach. However it comes out, the bet stops being a matter of taste, which is what this debate has needed most.
The intuition behind the bet is not mysterious. A fact about the user, an event that happened last Tuesday, and a procedure for handling a recurring task have different shapes. They differ in how they should be keyed, how they age, what supersedes them, and what a correct retrieval even looks like. A fact wants to be found by meaning. An event wants to be found by time and context. A procedure wants to be found by task shape, and critically, wants to be found before the work starts, not during. Uniform pipelines flatten those differences, and the flattening is precisely where quality leaks.
PISA, another submission in this crop, approaches the same question from psychology rather than benchmarking. It is a pragmatic psych-inspired unified memory system that combines schema creation, updating, and evolution with symbolic reasoning and neural retrieval. What I take from it is not any specific mechanism but the direction of travel: the interesting psych-inspired designs are the ones where the cognitive structure has operational consequences, where being a schema means something happens to you that does not happen to a raw string.
And then there is my favorite entry in this cluster, from a different field entirely. Unraveling the Complexity of Memory in RL Agents: an Approach for Classification and Evaluation, another ICLR 2026 submission, comes out of the reinforcement learning and POMDP tradition, and it brings something the LLM memory conversation badly needs: precise long-term versus short-term and declarative versus procedural terminology, with evaluation methodology attached, not vibes. When a community with decades of formal rigor independently lands on a typed decomposition of agent memory, the "types are marketing" position starts to require explaining away an inconvenient convergence.
The Convergence Argument
That convergence is, to me, the strongest single point. The taxonomy I use did not come from RL theory. It came from cognitive science by way of building products and watching where flat memory failed. The RL community arrived at overlapping distinctions from formal foundations. The benchmark-driven work arrives at them by measuring uniform against type-aware pipelines. Psychology arrived first by studying the one memory system we know works at scale, the human one, where the semantic, episodic, and procedural distinctions have been load-bearing since the 1970s.
These traditions do not settle the taxonomy between them, and I want to be careful not to claim they do. What they establish is narrower and still valuable: independent fields keep finding related joints worth testing, which justifies a harder comparison rather than a victory lap. The domain keeps hinting, from every direction anyone approaches it, that it has structure.
What Would Change My Mind
Authority without falsifiability is just confidence, so let me name the kill conditions.
I would simplify if a preregistered comparison found a uniform baseline matching a type-aware system, within a predeclared practical margin, across evaluations that stress type-differentiated behavior: temporal reasoning over episodic history, procedure reuse under task variation, contradiction handling across sessions. I would make the same change if the distinct-strategy comparisons now entering the literature failed to hold up under independent replication. What does not count as evidence against types is a flat baseline winning on workloads where boring retrieval dominates, because nobody claimed types help you fetch last week's meeting notes.
I keep a private list of results that would force that retreat, and the reading that produced this essay added entries to it. That is the posture I think this field deserves: build with conviction, publish the conviction, and keep the receipts for what would falsify it.
The minimalists are doing the field a service. Every baseline that embarrasses an over-built system is a tax on unearned complexity, and the tax is fair. I just think that when the evidence is laid end to end, from benchmarks to psychology to RL formalism, the types keep earning their keep, and they earn it the only way anything should, by surviving comparison. As of this year's reading, they are surviving it.